CLUSTSE: Stata module to estimate the statistical significance of parameters when the data is clustered with a small number of clusters

clustse gives researchers several options for estimating the statistical significance of parameters when the data is clustered with a small number of clusters. The default method of the program is the cluster adjusted t-statistics (CATs), which are described in Ibgragimov & Muller (2010). This procedure runs the model inside of each cluster and uses the standard error of the parameter estimates across the clusters to make statistical inference. The program also gives the option to perform the pairs cluster bootstrap-t procedure, as described by Cameron, Gelbach, & Miller (2008) and implemented by the program "clusterbs". Finally, the program gives the option for linear models to perform the wild cluster bootstrap-t procedure, as described by Cameron, Gelbach, & Miller (2008) and implemented by the program "cgmwildboot" created by Judson Caskey. Esarey & Menger (forthcoming) show that these procedures provide adequate power and desirable false rejection rates for performing statistical inference, even for data with small numbers of clusters or clusters of uneven size. clustse implements fixed effects options for the cluster-adjusted t-statistics procedure, and allows the clustse program to call the updated clusterbs program appropriately as an option. The program is updated to handle omitted variables and failed models in a way that matches the sister program for R. Now, the program will retain cluster estimates with omitted variables but omit the affected variables from the reported results under the default options. Under the option force(yes), the program will drop cluster estimates for all variables if any one variable is omitted from the model. This update also addresses an error in the calculation of t-statistics caused by omitted variables. Furthermore, the program is updated to return more descriptive error messages if the procedure fails or if needed programs are not installed. The newest version of clustse allows for the user to specify the truncate option. If truncate(yes) is specified, the program will drop outlying clusters' coefficient estimates from the analysis. An outlying coefficient is defined as an estimate that is more than 6 times the inter-quartile range of the mean of all the coefficient estimates. This option allows for researchers to eliminate cluster estimates from the analysis when they do not fall within the assumed normal distribution of the parameter's mean estimate. This option only applies to the CATs procedure (when the method() option is not specified).